Chemometrics

• Computationally intensive • Multivariate (many variables) statistical analysis • Applied to chemical systems or processes

• Reduce complex datasets • Identify and quantify sample groupings • Optimise experimental parameters • Isolate important variables and identify covariance • Provide reproducible measures of data • Allow for better visualisation of data

Singular variable

Too simplistic approach for complex data * wouldnt be able to see an outlier when using univariate

Used to explore relationships between different variables to look for patterns

• Design of Experiments • Exploratory Data Analysis • Classification • Regression

• Used to work out which collection method might be best • Relates to experimental setup • Will affect evidence collection, storage, instrument selection, parameter optimisation

• Based on y = mx + c linear relationship • Maps the effect of multiple independent variables (predictors) upon dependent variables (respone) • Allows prediction of quantitative sample properties

• Dimensionality reduction • Pattern recognition technique - identify grouping • Visualise trends that may otherwise have gone unoticed • Determination of sample similarity in complex data

Exploring the data without any prior assumptions or knowledge of the samples

Building classification rules for known sample grouping (from EDA)

• Cluster Analysis (CA) • Principal Component Analysis (PCA)

• Unsupervised technique • Samples grouped into clusters based on calculated distance (measure of their similarity) • Either agglomerative or hierachical • Output is a dendrogram • Good intial technique - simplifies complex data • Not limited to quantitative data • Visualisation of relationships • Can tell you that there are groupings but not why

Taking individual samples and grouping them together to form clusters

Opposite of agglomerative, taking a cluster and filter down into individual samples

• Principal Component Analysis • Unsupervised technique • Assesses all variables and desides which are relevant then determines which variables are correlated

• When the algorithm finds a correlation between the data • The PC that describes the largest variation between the samples will be given the highest priority (PC1)

Usually after EDA

• If the first PC is not sufficient to describe the spread of data, the calculation is repeated to find PC2 • The process is continused until all the variability within the dataset has been accounted for and modelled

• The distance along each PC from the mean gives a sample its ‘score’ • Each sample will have a different score for each PC

The optimum number of PCs - majority of the information is in the first few PCs

• Is a map of the samples • Each point is one sample • Similar samples cluster

• Increased visualisation • Plot differing combinations of PCs for enhanced discrimination

• The link between the scores plot and the chemistry of the samples • Whereas a scores plot is a map of samples, a loadings plot is the map of variables - why PCA is better than CA • Tells you why samples are grouping • Can either be +ve or -ve • Values indicate weightings for a particular variable • Quicker to identify variables using PCA • Objective